Prompt vector generation method and apparatus, device, and medium
By acquiring multimodal information of target users, performing feature extraction and word embedding processing, and combining correlation analysis and information fusion techniques to generate target prompt vectors, the problem of personalization in existing prompt vector generation schemes is solved, thus improving the customization effect of financial services.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- PING AN TECH (SHENZHEN) CO LTD
- Filing Date
- 2026-01-14
- Publication Date
- 2026-05-29
AI Technical Summary
Existing prompt vector generation schemes are difficult to personalize and customize, making it difficult to provide personalized services to users in financial scenarios.
By acquiring multimodal information of target users, feature extraction and word embedding are performed to generate personalized query vectors. Based on relevance analysis and information fusion technology, relevant information is fused with a pre-built industry knowledge hint library to generate target hint vectors, combining user-specific features and global common features.
It enables personalized and customized processing of prompt vectors, improving the execution effect of downstream tasks such as insurance recommendations and risk control decisions.
Smart Images
Figure CN122114141A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of large model technology, and in particular to a method, apparatus, device and medium for generating cue vectors. Background Technology
[0002] As digitalization accelerates across industries, many business scenarios leverage large-scale modeling technology to provide customized services to customers. For example, in diverse scenarios such as intelligent customer service response, risk control decision-making, and claims review, corresponding financial services are offered based on customer needs. Cue vectors, as the core medium connecting customer needs with specific service tasks, can guide the task-oriented large-scale model to provide the necessary services to the customer.
[0003] Existing suggestion vector generation solutions mostly use uniform rule templates to generate suggestion vectors. The generic suggestion vectors generated by this approach are difficult to accurately represent the personalized needs of different customers. In financial scenarios, this makes it difficult to provide personalized services to users.
[0004] Therefore, there is an urgent need to propose a new method for generating prompt vectors to achieve customization and personalization of prompt vectors, thereby improving the performance of downstream tasks such as insurance recommendations and risk control decisions. Summary of the Invention
[0005] This invention provides a method, apparatus, device, and medium for generating prompt vectors, so as to achieve customization and personalization of prompt vectors, thereby improving the execution effect of downstream tasks.
[0006] Firstly, a method for generating cue vectors is provided, including: Obtain multimodal information of the target user, including information associated with downstream tasks; Feature extraction is performed on the multimodal information, and word embedding processing is applied to the extracted features to generate a personalized query vector for the target user. Based on correlation analysis and information fusion technology, relevant information is fused between the personalized query vector and the pre-built industry knowledge prompt library to generate the target prompt vector for the target user. The industry knowledge prompt library includes vector groups corresponding to each global user common feature under the target business scenario of the downstream task. The target hint vector is applied to the downstream task so that the downstream task generates personalized execution results for the target user based on the target hint vector.
[0007] Secondly, a cue vector generation device is provided, comprising: The acquisition module is used to acquire multimodal information of the target user, including information associated with downstream tasks; The first generation module is used to extract features from the multimodal information and perform word embedding processing on the extracted features to generate a personalized query vector for the target user. The second generation module is used to fuse relevant information between the personalized query vector and the pre-built industry knowledge prompt library based on correlation analysis and information fusion technology to generate the target prompt vector for the target user. The industry knowledge prompt library includes a vector group corresponding to each global user common feature in the target business scenario to which the downstream task belongs. The target hint vector is applied to the downstream task so that the downstream task generates personalized execution results for the target user based on the target hint vector.
[0008] Thirdly, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the prompt vector generation method provided in the first aspect.
[0009] Fourthly, a computer-readable storage medium is provided, which stores a computer program that, when executed by a processor, implements the steps of the prompt vector generation method provided in the first aspect.
[0010] The aforementioned suggestion vector generation method, apparatus, device, and medium first acquire the multimodal information of the target user, including information related to downstream tasks. Then, feature extraction is performed on the multimodal information, and word embedding processing is applied to the extracted features to generate a personalized query vector for the target user. Therefore, this personalized query vector carries the target user's personalized characteristics. Next, based on relevance analysis and information fusion techniques, relevant information is fused between the personalized query vector and a pre-built industry knowledge suggestion library to generate the target suggestion vector for the target user. The industry knowledge suggestion library includes vector groups corresponding to each global user common feature within the target business scenario of the downstream task. Thus, the target suggestion vector is determined by combining the target user's personalized characteristics and global user common features, achieving personalized and customized processing of the suggestion vector. When the downstream task generates personalized execution results for the user based on the target suggestion vector, it can simultaneously refer to the target user's personalized characteristics and global user common features, focusing on the target user's personalization while not deviating from the gradually emerging patterns or trends within the target business scenario, thereby improving the execution effect of the downstream task. Attached Figure Description
[0011] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments of the present invention will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0012] Figure 1 This is a schematic diagram of an application environment for a prompt vector generation method according to an embodiment of the present invention; Figure 2 This is a flowchart illustrating a prompt vector generation method according to an embodiment of the present invention; Figure 3 This is a flowchart illustrating a prompt vector generation method in another embodiment of the present invention; Figure 4 This is a flowchart illustrating a prompt vector generation method in another embodiment of the present invention; Figure 5 This is a schematic diagram of a prompt vector generation device in another embodiment of the present invention; Figure 6 This is a schematic diagram of the structure of a computer device according to an embodiment of the present invention. Detailed Implementation
[0013] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0014] This invention provides a method for generating cue vectors, which can be applied to applications such as... Figure 1 In this application environment, the client communicates with the server via a network. The server can obtain multimodal information about the target user from the client, then extract features from the multimodal information and perform word embedding processing on the extracted features to generate a personalized query vector for the target user. Based on relevance analysis and information fusion techniques, the server fuses the personalized query vector with a pre-built industry knowledge hint base to generate a target hint vector for the target user. This generated target hint vector can be applied to downstream tasks on the server, such as insurance recommendation tasks or risk control prediction in the financial field. Alternatively, the target hint vector can be sent to the client, which can then pass it to downstream tasks. It should be noted that this is merely an illustrative example of the interaction process between the server and client, and is not intended to limit it.
[0015] The client can include, but is not limited to, various personal computers, laptops, smartphones, tablets, and portable wearable devices. The server can be implemented using a standalone server or a server cluster consisting of multiple servers. The invention will now be described in detail through specific embodiments.
[0016] Please see Figure 2 As shown, Figure 2 A flowchart illustrating the prompt vector generation method provided in this embodiment of the invention includes the following steps: S201, Obtain multimodal information of the target user, including information related to downstream tasks.
[0017] For example, a downstream task can be understood as a task that needs to be executed with the help of a cue vector, that is, a task executed after the cue vector is obtained.
[0018] For example, in a financial context, downstream tasks could include personalized financial product recommendations, identification of financial fraud risks, and risk assessment of personal consumer loans, among others.
[0019] Target users can be understood as the service recipients of downstream tasks. For example, if the downstream task is an insurance recommendation task, then the target users could be users who log in to the financial insurance service platform.
[0020] Multimodal information refers to user information presented in multiple data formats. For example, in financial scenarios, there is text-based information (e.g., personal information forms), image-based information (e.g., user-uploaded ID card photos), and time-series information (e.g., browsing history of financial products), etc.
[0021] The acquired multimodal information includes information related to downstream tasks.
[0022] Based on the task objectives of downstream tasks, the information that needs to be acquired and is related to downstream tasks can be determined in advance.
[0023] For example, if the downstream task is personal consumer loan assessment, and its objective is to determine a user's repayment ability and default risk, then to assess repayment ability, the downstream task requires characteristics such as income level and job stability; and to assess default risk, it requires characteristics such as historical repayment records and loan delinquency status. Therefore, when determining the information associated with the downstream task, occupational information, income verification, and cash flow statements can be categorized as relevant. In this case, the multimodal information obtained from the target user could include photos of their income verification documents, job descriptions, and cash flow statements for the past six months.
[0024] S202 extracts features from multimodal information and performs word embedding on the extracted features to generate personalized query vectors for the target user.
[0025] For example, multimodal information is extracted into single-modal features according to modality category, and word embedding is performed on the extracted features to obtain feature vectors corresponding to different modalities. Then, the feature vectors corresponding to different modalities are linearly transformed, and all feature vectors after linear transformation are fused to obtain personalized query vectors.
[0026] For example, multimodal information may include form information filled out by the user, photos of the user's ID card or invoice uploaded by the user, and timing data of onboard sensors required for car insurance, etc.
[0027] Input the form information into the pre-trained language model, extract the output of the last layer of the pre-trained language model as the initial text vector, that is, obtain the feature vector corresponding to the text class; Input the user's uploaded ID card photo or invoice photo into the visual embedding model to obtain the initial image vector, that is, to obtain the feature vector corresponding to the image class; The time series data is input into the time series coding model, and the dynamic features of the time series data are extracted as the time series initial vector, that is, the feature vector of the time series class is obtained. A linear projection layer is used to perform a linear transformation on the feature vectors corresponding to different modalities, ensuring that the feature vectors for each modality maintain the same dimension. Then, all the linearly transformed feature vectors are concatenated and fused to obtain a high-dimensional fused vector. Further linear transformations can be performed on the high-dimensional fused vector to adjust its dimension, ultimately yielding a personalized query vector.
[0028] S203, based on correlation analysis and information fusion technology, integrates relevant information between the personalized query vector and the pre-built industry knowledge hint library to generate a target hint vector for the target user. The industry knowledge hint library includes vector groups corresponding to each global user common feature in the target business scenario of the downstream task. The target hint vector is applied to the downstream task so that the downstream task generates personalized execution results for the target user according to the target hint vector.
[0029] For example, global user common characteristics can be understood as generalized feature patterns summarized for a certain group of people (rather than individual users). For instance, these patterns can be extracted from a large amount of user data and can reflect the common characteristics, needs, or adaptation strategies of a certain group of people.
[0030] In one example, common global user characteristics may include the correlation patterns between user characteristics and business elements within the target business scenario. For instance, in a financial scenario, this could include the correlation patterns between age and insurance type preference, income level and insurance demand level, disease risk factors and claim probability, occupation type and insurance risk level, and insurance duration and renewal intention.
[0031] For example, in a financial scenario, the industry knowledge hint library could include a vector group corresponding to "the repayment ability of users with a monthly income of more than 20,000" and a vector group corresponding to "the default risk of users with mortgages".
[0032] When generating the target suggestion vector, the correlation between the personalized query vector and each vector group can be matched according to the degree of correlation between the personalized query vector of the target user and the vector groups corresponding to the common features of each global user in the industry knowledge suggestion library. This determines the degree of association between the personalized query vector and the vector groups corresponding to each common feature of the global user. Then, according to the degree of association, the vector groups with high degree of association are merged with the personalized query vector to obtain the target suggestion vector. In this way, information related to the personalized query vector of the target user in the industry knowledge suggestion library can be integrated into the target suggestion vector.
[0033] The correlation analysis between vectors can be achieved by calculating the scaling click or cosine similarity between vectors, or by other correlation analysis techniques, which are not limited in this application.
[0034] Downstream tasks can generate personalized execution results for target users based on the target cue vector, such as providing personalized insurance recommendations or predicting the credit risk of the target user.
[0035] In summary, the suggestion vector generation method provided in this application first obtains the multimodal information of the target user, including information related to downstream tasks. Then, it extracts features from the multimodal information and performs word embedding on the extracted features to generate a personalized query vector for the target user. Therefore, this personalized query vector carries the target user's personalized characteristics. Next, based on relevance analysis and information fusion techniques, it fuses the personalized query vector with a pre-built industry knowledge suggestion library to generate the target suggestion vector for the target user. The industry knowledge suggestion library includes vector groups corresponding to each global user common feature within the target business scenario of the downstream task. Thus, the target suggestion vector is determined by combining the target user's personalized characteristics and global user common features, achieving personalized and customized processing of the suggestion vector. When the downstream task generates personalized execution results for the user based on the target suggestion vector, it can simultaneously refer to the target user's personalized characteristics and global user common features. While focusing on the target user's personalization, it does not deviate from the gradually emerging patterns or trends within the target business scenario, thereby improving the execution effect of the downstream task.
[0036] In some embodiments, the vector group corresponding to the global user common features includes a global cue vector, a key vector corresponding to the global cue vector, and a value vector corresponding to the global cue vector. For example... Figure 3 As shown, the vector group corresponding to the common features of all users can be obtained in the following ways: S301 extracts global user common features from the big data of the target business scenario.
[0037] For example, one can first filter valid data for the target business scenario from a large database, such as credit application and repayment data for the past three years in the consumer loan scenario.
[0038] Then, this data is cleaned. In one example, cleaning can be done by removing missing and outlier values. For instance, removing fake income statements from financial data.
[0039] Then, the cleaned data is grouped. For example, in a financial insurance scenario, the data is grouped by fields such as age, gender, income level, occupation, repayment history, etc.
[0040] After grouping, the data is standardized.
[0041] Extract group patterns from the grouped data to form candidate features. For example, analyze the commonalities in repayment periods (such as 90% of users choosing 12-24 months) and default rates of "consumer loan users with a monthly income of over 20,000 yuan", as well as commonalities in asset allocation (such as 60% of users having wealth management products).
[0042] The extraction of group patterns can be achieved through statistical analysis or clustering algorithms. These will not be elaborated upon here.
[0043] Finally, the candidate features are validated, and those that pass the validation are used as common features for all users. For example, this might yield the relationship between age and insurance needs, income level and insurance needs, and so on.
[0044] S302, based on large model word embedding technology, performs word embedding processing on global user common features to obtain global prompt vector, key vector and value vector.
[0045] Among them, the global prompt vector is used to carry the core semantics of common features of global users.
[0046] The key vector is the identifier vector of the global prompt vector, used to carry the keyword semantics in the global user common features; The value vector is used to carry the complete value semantics of the common features of global users.
[0047] For example, if the common characteristics of global users are the text "30-40 years old, the breadwinner of the family, priority is given to critical illness insurance with a coverage of 500,000 yuan, supplemented by accident insurance, and the premium budget is 5%-8% of the annual income".
[0048] Core prompts can be extracted from the text, such as the critical illness insurance and accident insurance needs of family breadwinners aged 30-40. These prompts are then transformed into vectors using large-scale model word embedding technology fine-tuned by the insurance industry, resulting in a global prompt vector.
[0049] The core prompts are further reduced to search keywords, such as "30-40 years old," "family breadwinner," "critical illness," and "accident insurance." These are also transformed into vectors through large-scale model word embedding, resulting in key vectors. Compared to global prompt vectors, key vectors reduce redundant semantic interference, improving matching efficiency when subsequently matching relevance with personalized query vectors.
[0050] The complete value information of users' common characteristics, such as prioritizing critical illness insurance with a coverage of 500,000 yuan, supplemented by accident insurance, and budgeting premiums of 5%-8% of annual income, is processed by word embedding to obtain a value vector.
[0051] In one example, vector groups can also be generated in the following way: Regardless of whether the common features of global users belong to text, image, time series, or other modalities, a unified semantic vector can be generated first through large model word embedding technology. The unified semantic vector serves as the common semantic source for the global prompt vector, key vector, and value vector to ensure semantic consistency among the three. Then, through three independent learnable linear transformation matrices, the unified semantic vector is transformed into global prompt vector, key vector, and value vector that perform different functions, respectively.
[0052] Specifically, the vector generation logic, based on the Transformer attention mechanism, configures three independent learnable linear transformation matrices for the unified semantic vector. The linear transformation matrix used to generate the global cue vector is used to extract the core semantics of common features from the unified semantic vector. The linear transformation matrix used to generate the key vector is used to extract the retrieval semantics for matching from the unified semantic vector. The linear transformation matrix used to generate the value vector is used to extract the complete business value semantics from the unified semantic vector.
[0053] The linear transformation matrix can be optimized through downstream task training. For example, a vector set is generated using the current linear transformation matrix, and a target cue vector is generated based on this vector set. A downstream insurance recommendation task is then performed based on the target cue vector. The execution result of the downstream insurance recommendation task is compared with the label result to determine the loss value. The parameters of the linear transformation matrix are adjusted by backpropagating the loss value. Through repeated training in this way, the final linear transformation matrix is obtained.
[0054] In this embodiment, global user common features are extracted from a large database of business scenarios. These global user common features are then transformed into machine-processable vectors. All vector groups of global user common features are mapped into the same large model word embedding space, allowing the correlation between each common feature to be determined based on the parameter values of the vector groups. When transforming global user common features into vectors, this application generates semantically hierarchical vector groups, enabling different vectors in each group to perform different functions, facilitating the application of the industry knowledge hint library in subsequent processing. For example, key vectors with lower semantic redundancy are used in matching, retrieval, and preliminary similarity calculation processes, thereby reducing computational load.
[0055] In some embodiments, such as Figure 4 As shown, step S203 above, "based on relevance analysis and information fusion technology, performs relevant information fusion on the personalized query vector and the pre-built industry knowledge hint base to generate the target hint vector for the target user," includes the following steps: S401, calculate the correlation between the personalized query vector and each key vector in the industry knowledge hint library based on the attention mechanism, and obtain the attention weight corresponding to each key vector.
[0056] For example, based on the attention mechanism, the personalized query vector of the target user (representing user characteristics) is matched with each key vector in the industry knowledge hint library (representing a certain industry consensus or common feature), and the semantic similarity between the two is calculated. This similarity is the attention weight of the corresponding key vector. The higher the weight, the more relevant the industry consensus is to the user.
[0057] For example, in the medical insurance recommendation scenario, the personalized query vector is the vector of user A (a 35-year-old hypertensive patient with an average daily sodium intake of 4g and exercise once a week); the key vector of the industry knowledge prompt library contains multiple key vectors such as "sodium excess characteristics of hypertensive users aged 30-40" and "sedentary characteristics of office workers aged 25-35". Using attention mechanism formulas (e.g., the scaling dot product formula), the similarity between user A's personalized query vector and each key vector is calculated. For example, the similarity (weight) between the personalized query vector and the key vector "characteristics of sodium excess in 30-40 year old hypertensive users" is 0.8, and the similarity (weight) between the key vector and the key vector "characteristics of sedentary office workers aged 25-35" is 0.2. This allows for improved information fusion by focusing on information in the industry knowledge base that is highly relevant to the personalized query vector during subsequent information fusion processes.
[0058] S402, based on attention weights, performs a weighted summation of the value vectors corresponding to each key vector to obtain a personalized correction vector.
[0059] For example, the attention weight corresponding to each key vector is multiplied by the value vector corresponding to that key vector, and then all the vectors obtained after the multiplication are summed to obtain the personalized correction vector. In this way, the personalized correction vector carries the business value semantics of the global user common features that are most relevant to the target user.
[0060] S403, based on attention weights, performs a weighted summation of the global cue vectors corresponding to each key vector to obtain the target global cue vector.
[0061] For example, the attention weight corresponding to each key vector is multiplied by the global cue vector corresponding to that key vector, and then all the vectors obtained from the multiplication are summed to obtain the target global cue vector. In this way, the target global cue vector carries the core semantics of the common features of the global users that are most relevant to the target user.
[0062] S404 merges the personalized correction vector and the target global cue vector to generate the target cue vector.
[0063] For example, the personalized correction vector and the target global cue vector can be superimposed to generate the target cue vector.
[0064] In some embodiments, step S404, "fusing the personalized correction vector and the target global cue vector to generate the target cue vector", may include the following steps: determining the fusion coefficient based on the attention weights corresponding to all key vectors; For example, the average of the attention weights corresponding to all key-value vectors is calculated as the fusion coefficient.
[0065] Alternatively, the fusion coefficients can be obtained by performing a nonlinear transformation on the average value, where the parameters of the nonlinear transformation can be obtained through training with a large number of samples.
[0066] After determining the fusion coefficients, the target cue vector can be determined using the following formula: ; in, Provide a target hint vector. Provide a global hint vector for the target. For personalized correction vectors, This is the fusion coefficient.
[0067] In this embodiment, the correlation between the target user's personalized query vector and the key vectors corresponding to various global user common features in the industry knowledge base is calculated to determine the correlation between the target user's personalized features and various global user common features. Then, according to the attention weight corresponding to the degree of correlation, the global prompt vectors corresponding to various global user common features are weighted and summed, and the value vectors corresponding to various global user common features are also weighted and summed. The target prompt vector and the personalized correction vector obtained after weighted summation are fused to obtain the target prompt vector. The target prompt vector increases the attention to industry consensus related to the target user, thereby avoiding deviation from industry consensus when using the target prompt vector to execute downstream tasks. This ensures that the task execution results conform to industry rules while paying attention to user personalized features, making the structure more effective.
[0068] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.
[0069] In one embodiment, a prompt vector generation device is provided, which corresponds one-to-one with the prompt vector generation method in the above embodiments. For example... Figure 5 As shown, the prompt vector generation device 500 includes: an acquisition module 501, a first generation module 502, and a second generation module 503. Detailed descriptions of each functional module are as follows: The acquisition module 501 is used to acquire multimodal information of the target user, the multimodal information including information associated with downstream tasks; The first generation module 502 is used to extract features from the multimodal information and perform word embedding processing on the extracted features to generate a personalized query vector for the target user. The second generation module 503 is used to perform relevant information fusion on the personalized query vector and the pre-built industry knowledge prompt library based on correlation analysis and information fusion technology to generate the target prompt vector for the target user. The industry knowledge prompt library includes a vector group corresponding to each global user common feature in the target business scenario to which the downstream task belongs. The target hint vector is applied to the downstream task so that the downstream task generates personalized execution results for the target user based on the target hint vector.
[0070] In some embodiments, the vector group corresponding to the global user common features includes a global hint vector, a key vector corresponding to the global hint vector, and a value vector corresponding to the global hint vector; the device 500 further includes a construction module, configured to obtain the vector group corresponding to the global user common features in the following manner: Extract the common global user features from the large database of the target business scenario; Based on the large model word embedding technology, word embedding processing is performed on the global user common features to obtain the global prompt vector, the key vector and the value vector; The global suggestion vector is used to carry the core semantics of the common features of global users; The key vector is the identifier vector of the global prompt vector, used to carry the keyword semantics in the global user common features; The value vector is used to carry the complete value semantics of the common features of global users.
[0071] In some embodiments, the second generation module 503 is specifically used for: The correlation between the personalized query vector and each key vector in the industry knowledge hint base is calculated based on the attention mechanism to obtain the attention weight corresponding to each key vector; Based on the attention weights, the value vectors corresponding to each key vector are weighted and summed to obtain the personalized correction vector. Based on the attention weights, the global cue vectors corresponding to each key vector are weighted and summed to obtain the target global cue vector. The personalized correction vector and the target global prompt vector are fused to generate the target prompt vector.
[0072] In some embodiments, the second generation module 503 is specifically used for: The fusion coefficients are determined based on the attention weights corresponding to all the key vectors. The target cue vector is determined using the following formula: ; in, The target cue vector, The target global cue vector, the For personalized correction vectors, This is the fusion coefficient.
[0073] In some embodiments, the second generation module is specifically used for: The average value of the attention weights corresponding to all the key vectors is calculated as the fusion coefficient.
[0074] In some embodiments, the first generation module 502 is specifically used for: The multimodal information is processed by extracting single-modal features according to modality category, and word embedding is performed on the extracted single-modal features to obtain feature vectors corresponding to different modalities. The feature vectors corresponding to different modalities are linearly transformed, and all the feature vectors after linear transformation are fused to obtain the personalized query vector.
[0075] In the above embodiments, the global user common characteristics include the correlation patterns between user characteristics and business elements under the target business scenario.
[0076] In summary, the prompt vector generation device provided in this application first acquires the multimodal information of the target user, including information related to downstream tasks. Then, it extracts features from the multimodal information and performs word embedding processing on the extracted features to generate a personalized query vector for the target user. Therefore, this personalized query vector carries the target user's personalized features. Next, based on relevance analysis and information fusion techniques, it fuses the personalized query vector with a pre-built industry knowledge prompt library to generate a target prompt vector for the target user. The industry knowledge prompt library includes vector groups corresponding to each global user common feature within the target business scenario of the downstream task. Therefore, it can be seen that the target prompt vector is determined by combining the target user's personalized features and the global user common features, achieving personalized and customized processing of the prompt vector. When the downstream task generates personalized execution results for the user based on the target prompt vector, it can simultaneously refer to the target user's personalized features and the global user common features. While focusing on the target user's personalization, it does not deviate from the gradually emerging patterns or trends within the target business scenario, thereby improving the execution effect of the downstream task.
[0077] Specific limitations regarding the cue vector generation device can be found in the limitations of the cue vector generation method described above, and will not be repeated here. Each module in the aforementioned cue vector generation device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module.
[0078] In one embodiment, a computer device is provided, such as Figure 6 As shown, the computer device 600 includes a memory 601, a processor 602, and a computer program stored in the memory 601 and executable on the processor 602. When the processor 602 executes the computer program, it implements the steps of the prompt vector generation method of any of the above embodiments.
[0079] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps of the prompt vector generation method of any of the above embodiments.
[0080] It should be noted that the functions or steps that can be implemented by the computer-readable storage medium or computer device described above can be referred to the description of each step of the data risk prediction method based on artificial intelligence in the foregoing method embodiments. To avoid repetition, they will not be described one by one here.
[0081] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.
[0082] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.
[0083] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.
Claims
1. A method for generating cue vectors, characterized in that, The method includes: Obtain multimodal information of the target user, including information associated with downstream tasks; Feature extraction is performed on the multimodal information, and word embedding processing is applied to the extracted features to generate a personalized query vector for the target user. Based on correlation analysis and information fusion technology, relevant information is fused between the personalized query vector and the pre-built industry knowledge prompt library to generate the target prompt vector for the target user. The industry knowledge prompt library includes vector groups corresponding to each global user common feature under the target business scenario of the downstream task. The target hint vector is applied to the downstream task so that the downstream task generates personalized execution results for the target user based on the target hint vector.
2. The method according to claim 1, characterized in that, The vector group corresponding to the global user common features includes a global suggestion vector, a key vector corresponding to the global suggestion vector, and a value vector corresponding to the global suggestion vector; the vector group corresponding to the global user common features is obtained in the following way: Extract the common global user features from the large database of the target business scenario; Based on the large model word embedding technology, word embedding processing is performed on the global user common features to obtain the global prompt vector, the key vector and the value vector; The global suggestion vector is used to carry the core semantics of the common features of global users; The key vector is the identifier vector of the global prompt vector, used to carry the keyword semantics in the global user common features; The value vector is used to carry the complete value semantics of the common features of global users.
3. The method according to claim 2, characterized in that, The method, based on relevance analysis and information fusion technology, fuses relevant information between the personalized query vector and a pre-built industry knowledge hint base to generate a target hint vector for the target user, including: Based on the attention mechanism, the correlation between the personalized query vector and each key vector in the industry knowledge hint library is calculated to obtain the attention weight corresponding to each key vector; Based on the attention weights, the value vectors corresponding to each key vector are weighted and summed to obtain the personalized correction vector. Based on the attention weights, the global cue vectors corresponding to each key vector are weighted and summed to obtain the target global cue vector. The personalized correction vector and the target global prompt vector are fused to generate the target prompt vector.
4. The method according to claim 3, characterized in that, The step of fusing the personalized correction vector and the target global cue vector to generate the target cue vector includes: The fusion coefficients are determined based on the attention weights corresponding to all the key vectors. The target cue vector is determined using the following formula: ； in, The target cue vector, The target global cue vector, the For the personalized correction vector, The fusion coefficient is denoted as .
5. The method according to claim 4, characterized in that, The step of determining the fusion coefficient based on the attention weights corresponding to all the key vectors includes: The average value of the attention weights corresponding to all the key vectors is calculated as the fusion coefficient.
6. The method according to any one of claims 1-5, characterized in that, The step of extracting features from the multimodal information and performing word embedding processing on the extracted features to generate a personalized query vector for the target user includes: The multimodal information is processed by extracting single-modal features according to modality category, and word embedding is performed on the extracted single-modal features to obtain feature vectors corresponding to different modalities. The feature vectors corresponding to different modalities are linearly transformed, and all the feature vectors after linear transformation are fused to obtain the personalized query vector.
7. The method according to claim 1, characterized in that, The global common user characteristics include the correlation patterns between user characteristics and business elements in the target business scenario.
8. A cue vector generation device, characterized in that, The device includes: The acquisition module is used to acquire multimodal information of the target user, including information associated with downstream tasks; The first generation module is used to extract features from the multimodal information and perform word embedding processing on the extracted features to generate a personalized query vector for the target user. The second generation module is used to fuse relevant information between the personalized query vector and the pre-built industry knowledge prompt library based on correlation analysis and information fusion technology to generate the target prompt vector for the target user. The industry knowledge prompt library includes a vector group corresponding to each global user common feature in the target business scenario to which the downstream task belongs. The target hint vector is applied to the downstream task so that the downstream task generates personalized execution results for the target user based on the target hint vector.
9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the prompt vector generation method as described in any one of claims 1 to 7.
10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the prompt vector generation method as described in any one of claims 1 to 7.